One source of truth for every AI tool.
Today, Forge is a dead-simple sync engine — point it at a repo, it distributes skills to Claude Code, Cursor, Windsurf, Copilot, and Cline in the right format. But that's just the foundation. Forge is becoming the Harness Compiler — same CLI, new brain: dependency graphs, tree-shaking, token budgets, per-agent optimization. Think Vite, but for AI agents.
Two minutes to set up. Zero maintenance.
Global setup (once per machine)
Run forge init --global and point it at your team's shared resource repo. The wizard detects your AI tools and writes the config. That's it — every new project you open gets synced automatically.
$ forge init --global
✓ Done. Forge will sync automatically.
Per-project setup (once per repo)
Need project-specific skills? Run forge init inside the repo. Same wizard, same simplicity. Once configured, forget about it — Forge keeps everything in sync.
$ forge init
✓ Project-level config created.
No YAML to hand-edit. No complex configs to learn. The wizard handles everything.
Best paired with an AI Resources Platform
Forge pulls from any git repo. But for the best experience, we recommend using a structured AI Resources Platform as your source — it gives you convention-over-configuration, versioned resources, automated testing, and a clear lifecycle from idea to delivery.
See it in action
Set & Forget
Run the wizard once. It detects your AI tools, adds a git source, enables auto-sync scheduling. Never think about it again.
Easy Updates
Check status, spot drift, add a second source, sync everything. No YAML editing, no guessing — the wizard handles it.
Power User
Fine-grained source control with subdirectories, include/exclude filters, merge strategies, dry-run preview, and health dashboard.
12 commands. Everything you need.
| Command | What it does |
|---|---|
| forge init | Interactive wizard — detects AI tools, configures sources, writes config. |
| forge sync | Fetches from all sources, resolves targets, places files for each assistant. |
| forge status | Shows drift between last sync and current filesystem state. |
| forge health | SLI dashboard — sync frequency, drift count, source reachability. |
| forge config | Interactive config editor with section-based menu. |
| forge source list | Shows configured sources with type, synced file counts. |
| forge source add | Add a new source interactively. |
| forge source remove | Remove a source with optional cleanup of synced files. |
| forge source edit | Edit existing source configuration with pre-filled values. |
| forge scheduler | Install/uninstall/status for OS-level auto-sync scheduling. |
| forge telemetry | Shows telemetry status, config, and what's collected. |
| forge uninstall | Removes .crucible directory, scheduler, and all synced files. |
5 assistants. One config.
| Assistant | Config Paths | Transforms |
|---|---|---|
| Claude Code | .claude/skills/, .claude/rules/, .claude/commands/ | None (passthrough) |
| Cursor | .cursor/rules/ | .md → .mdc + frontmatter injection |
| Windsurf | .windsurf/rules/, .windsurf/workflows/ | Frontmatter injection only |
| GitHub Copilot | .github/skills/, .github/instructions/ | .md → .instructions.md + frontmatter |
| Cline | .cline/skills/, .clinerules/ | None (passthrough) |
Adding a new assistant = one file. Zero command handler changes.Learn more →
Built for real teams
Multi-source
Parallel fetching via Promise.allSettled. Failed sources don't block others.
Conflict resolution
Source + local conflict strategies: last-wins, first-wins, or error.
Backup
When remote-wins overwrites local changes, .bak copies are created automatically.
SHA-256 detection
No unnecessary rewrites. Only writes when content actually changed.
Dry-run
Preview all changes without writing. See what would happen before it does.
Rollback
On mutation failure, already-written files are rolled back. No partial state.
Scheduler
Cross-platform auto-sync: launchd (macOS), crontab (Linux), schtasks (Windows).
Health dashboard
SLI metrics: sync frequency, drift count, source reachability, active assistants.
Three source layouts, auto-detected
Canonical
Flat files organized as skills/, rules/, workflows/ — distributed and transformed per assistant.
Pre-structured
Contains assistant-specific directories (.claude/, .cursor/, etc.) — copied as-is without transforms.
Hybrid
Mix of both — canonical files transformed and distributed; pre-structured files copied directly. Auto-detected by default.
Coming next: Forge becomes the Harness Compiler
Same forge CLI you already use — but under the hood, a full compiler pipeline. Other tools install files. Forge will compile your AI environment.
Dependency graph
Skills declare dependencies via frontmatter. Forge resolves the full DAG, detects cycles, and orders output correctly.
Tree-shaking
Java project? Python skills are automatically excluded. Forge detects project context from marker files and prunes irrelevant skills.
Per-agent optimization
Claude Code gets priority-ordered individual files. Cursor gets merged rule files. Each agent gets skills in its optimal format.
Token budgeting
Set per-agent context limits. When total exceeds budget, lowest-priority skills are dropped first. Dependencies are always protected.
Build manifest
Every compile emits a manifest — skill IDs, versions, hashes, agent targets, token estimates. This is Ember's data source.
Smithery integration
Pull skills from agentskills.io as a source. Their 25k-star skill ecosystem becomes compiler input.
Ready to get started?
Install Forge, run forge init, and sync your first config in under 2 minutes. The compiler features are coming in v0.7.